Bias correction of satellite radiances
Biases in satellite radiance observations with respect to radiances modelled from NWP backgrounds arise from several sources, including instrument characteristics, systematic forward model (radiative transfer) errors and biases generated in the forecast model that is used to produce the background. The biases may be large compared to the signals that assimilation systems aim to analyse, and bias correction schemes have been found to be necessary for the radiances to give benefit in operational NWP systems. Whilst forecast model errors are an important source of error, and their treatment in assimilation is an active area of research, current operational schemes correct observations only.
The documents collected here describe the systems implemented by each of the four contributing centres. All formulate biases in terms of predictors, which may contain model or observation information, with coefficients, and all have moved on from using purely static schemes, where bias statistics are derived retrospectively from timeseries of observed and forecast radiances. Three of the centres (ECMWF, Météo-France and the UK Met Office) now operate variational bias correction schemes with similar theoretical bases whereby, in each NWP cycle, the analysis allows a first-guess bias to vary as part of the control vector, with an appropriate constraining error term. DWD also operate an adaptive scheme, which is run in every cycle to produce a correction from previous statistics but with a weighting towards more recent cycles that allows for a similar concept of “adaptation rate” as is inherent in the other, variational, schemes. There are other similarities, the theoretical basis for which DWD discuss in their contribution.
Practicalities of implementation are also described. These include predictor choice, target adaptation rate and the use of certain channels without a bias correction to “anchor” the model where necessary. Where appropriate there is discussion of the special considerations required for limited-area models, in which observation counts are typically low and weather regimes may fluctuate more rapidly.
The NWP SAF near-real time (NRT) quality monitoring provides data quality information before and after bias correction, as well as bias information (in terms of corrections).